SentientEdge

Research letters · 19 October 2026

Research letter, 19 October 2026

This week

The strongest papers this fortnight are about how evidence fails, not about new signals. One shows that the Deflated Sharpe Ratio, the overfitting control most of us (SentientEdge included) rely on, can be made to approve pure noise once each new idea is chosen after seeing the last result. Another shows that strict admission gates mostly just reject everything. On the signal side, the cleanest ML result of the week comes from portfolio construction rather than the model: net Sharpe goes from −0.38 to +0.42 with no change in forecast quality. A carefully hindsight-free news-topic paper ends up with a monthly Sharpe of 1.12 that is no better than a price-only baseline. Regime detection was quiet. Nothing new had out-of-sample evidence strong enough for a full item, so that section is short.

A note on sourcing: for the three SSRN items, numbers come from the abstracts and author repositories, not full-text reads. This is flagged where it matters.

1. Deflate by Bits, Not Trials: Why Query-Counted Deflated Sharpe Ratios Fail Under Adaptive Search, and a Sealed-Holdout Remedy

Shlok Singh Sobti · posted 5 Oct 2026 · SSRN 7557458 · code: github.com/shloksobti/deflate-by-bits

Method. The DSR raises the significance bar with the number of trials N, and the current best practice is to log every backtest so that N is honest. Sobti argues that N measures the wrong thing whenever the researcher is adaptive, i.e. picks the next candidate after seeing earlier validation results. In a "combine what worked" procedure over k pure-noise strategies, the validation t-stat grows like √(2k/π), while the DSR hurdle grows only like √(2 ln k). So the DSR eventually certifies noise with probability approaching 1. His proposed fix is to deflate by the information the holdout reveals (log of the number of possible transcripts) rather than by query count. In practice that means a sealed holdout: an opaque object whose only operation is a budgeted PASS/FAIL verdict, so K verdicts leak at most about log K bits.

What they tested and found.

Limitations. This is mainly a statistics result. The real-data part tests whether the procedure stays valid, not whether any strategy has alpha. The "combine what worked" attack is an extreme case; a disciplined human researcher leaks much less than an agent running thousands of queries. The bit-counting bound is also only as good as your ability to bound the transcript set, and a person reading a full equity curve leaks far more than a PASS/FAIL. This summary rests on the abstract and the repository's documentation, not the full proofs.

Evidence vs speculation.

What we will do with it. Our August review of the production scorer reads like a textbook adaptive campaign: three narratives in sequence, each prompted by the previous result, all evaluated on the same window. So we are declaring the live shadow books formed from 26 October onward a sealed holdout for the question of clean momentum versus the frozen scorer. One PASS/FAIL statistic is pre-registered, median 60-day net excess vs SPY on matched formations, with a budget of three looks at 13, 26 and 52 weeks and no intermediate equity curves. If anyone needs an unbudgeted look before week 26, it is logged as a breach and the decision date moves back 26 weeks. Two breaches, and we drop the sealed-holdout idea as operationally infeasible for this team.

2. On the Boundary of Admission Gates: An Injected-Truth Study of Falsification-First Selection in Quantitative Strategy Research

Tianlun Zheng (Fudan) · 6 Oct 2026 · arXiv:2610.07701 · code + frozen data: github.com/simplify23/quant-trading-agent

Method. A three-arm design with a known injected signal:

What they tested and found. Synthetic panel, 400 repetitions per cell:

Training t (signal, ann.)Arm A FDRArm B adoptionArm B FDR
0.72 (10%)59.0%1.7%0%
1.44 (20%)24.3%4.7%0%
2.87 (40%)1.0%60.3%0%
5.74 (81%)0%100%0%

Limitations. The signal is a constant injected alpha, which is much friendlier than real decaying, regime-dependent alpha. The real-data panel is tiny (29 ETFs, 745 observations, one market). Only N = 40 trials were tested. Credit to the author for reporting that his first two control designs were invalid; that honesty raises trust in the rest.

Evidence vs speculation.

What we will do with it. Our promotion gates have only ever been seen killing things. We have never measured their recall. We will build 200 placebo books from random rankings through our normal book-construction machinery, inject a known excess return into each at t = 1.0, 1.5, 2.5 and 4.0, and run the production gate stack on each, once on excess-vs-SPY returns and once on absolute returns. If recall at t = 2.5 is below 50%, the gates are too strict to ever promote a realistic signal and get flagged for redesign, not loosened on the fly. If any gate gives materially different verdicts on absolute versus excess returns, it has this paper's bug and is fixed before the next promotion decision. If recall is at least 50% and false discoveries at t = 0 are at most 5%, the gates are validated and nothing changes.

3. Trading Costs and Machine-Learning Return Forecasts

Daniele Bianchi, Teng Jiao (QMUL) · posted 16 Oct 2026 · SSRN 7617338

Method. Standard US cross-sectional ML forecasts (elastic net, boosted trees, neural network) evaluated where capital can actually be deployed. They try three fixes: (i) reweighting the training loss toward liquid stocks; (ii) training only on liquid stocks; (iii) leaving the model alone but building the book differently, with liquidity-weighted holdings and trading only when the forecast covers the half bid-ask spread.

What they tested and found.

Limitations. This summary rests on the abstract, so the sample period, the cost model, and how the $500M price impact was calibrated are unchecked, and those choices decide the headline. A net Sharpe of 0.42 at scale is respectable but not impressive, and the confidence interval is probably wide. The half-spread filter is a well-known no-trade band, so the novelty is the clean decomposition, not the trick itself.

Evidence vs speculation.

What we will do with it. Our August review found that our production machinery doesn't tax away momentum, which was encouraging. The open question is turnover. We will add a shadow variant of the clean momentum candidate with a no-trade band: an incumbent is replaced only when the newcomer's rank gap implies an expected gain above that name's half-spread, or when the incumbent falls out of the top 50. After 26 live weeks, the variant is killed unless it cuts turnover by at least 30% and keeps median 20-day net excess vs SPY within 0.25 percentage points of the base candidate on matched formations. If it passes, it is a construction candidate for whichever signal eventually wins, not a promotion.

4. Reading the News Without Hindsight: NoLBERT Topics and the Cross-Section of Stock Returns

Kevin Foley, Jonathan Hartadi, Shivesh Prakash, Charishma Takkallapalli, Swapnil Vatsal · 15 Oct 2026 · arXiv:2610.22745 (also SSRN 7617420) · code and derived topics in project repo (CRSP rebuild script provided)

Method. The point is to remove lookahead in the text model. NoLBERT is pretrained only on text from 1976–1995, then made into a sentence encoder by contrastive fine-tuning on same-article passage pairs, with no labels and no returns. They compare six "readers" (LDA, MiniLM, FinBERT, NoLBERT variants) × 26 feature sets × 8 models (OLS through IPCA, XGBoost and EBM), about 600 experiments per frequency. Data: 1.26M articles from 2012–2024 on the 1,000 largest US stocks. Train 2013–19, validate 2020–21, test 2022–24, with costs of 10 bp per dollar traded.

What they tested and found.

Limitations. The test window is three years, which is short, and 2022 is a single dominant regime. The ensemble is equal-weighted over a search that was itself chosen with hindsight about which readers to include. The authors are admirably upfront that the headline doesn't beat the price benchmark.

Evidence vs speculation.

What we will do with it. This paper is a good template for a null benchmark, more than a signal source. For any future signal we test, earnings drift or any news feature, we will require an incremental comparison against a price-only ensemble (clean 12-1 momentum plus short-term reversal, equal-weighted ranks), not against zero or SPY. A candidate is dead if the paired test of candidate-plus-price versus price-only, on matched weekly formations, gives p > 0.10 at the 20-day horizon after 26 live weeks, however good its standalone Sharpe looks.

5. What Survives Honest Testing? A Pre-Registered Test of Eleven Trading Signals

Yasanji Ratnaike · posted 2–3 Oct 2026 · SSRN 7553278

Method. Eleven well-known signals (pairs/stat-arb, factor, and trend) in European equities and global futures. Rules were fixed before testing, with separate development, crisis and hold-out periods, plus multiple-testing corrections, factor attribution and a power analysis.

What they tested and found.

Limitations. Abstract-level read only: the cost model, the exact holdout dates, and how strict the pre-registration was (a timestamped public registry versus a declaration in the paper) are unverified. The universe is European, not US single names. Eleven signals is a small family, which is good for honesty but means the result doesn't say much about anything not tested.

Evidence vs speculation.

What we will do with it. The power point is the one that applies to us. Our own 60-day head-to-head between two candidates flipped sign between halves of an 82-week sample. As a desk calculation, we will take the observed dispersion of matched-formation excess-return differences and compute the minimum detectable effect at 80% power for 26, 52 and 104 live weeks, with overlap-adjusted effective sample size. If the minimum detectable effect at 52 weeks is above 4 percentage points, "pick the winner from the live shadow" is formally retired as the promotion method, and both signals become candidates for a blend or a regime-conditional switch, to be tested with its own placebo. If it is at most 4 points, the current plan stands and the decision date is fixed to that horizon now.

6. From Economic Explanation to Performance Improvement in Machine-Learning Portfolios

Yunqi Liang, Stefan Zohren, Mihai Cucuringu (Oxford) · 10 Oct 2026 · arXiv:2610.15144

Method. First-order attribution that maps an ML allocation function's feature exposures onto sensitivities of portfolio metrics (Sharpe), linked to Shapley values for portfolio objectives, with bootstrap inference. They then intervene post-training: shrink or boost feature exposures based on prior-period attributions, with "temporal stabilisation" (accumulating attributions over time) and cross-model averaging.

What they tested and found. Daily out-of-sample trading 2006–2024, with costs of 5 bp per unit turnover.

Limitations. No code link was found. Most of the gain coincides with turnover halving, so a good part of this may simply be "smooth your signal," which a plain EMA could match. That ablation is the one we would want to see, and it does not appear in the portion reviewed. Universe details were unclear. There are many design choices (stabilisation window, averaging) and no visible multiple-testing accounting for them.

Evidence vs speculation.

What we will do with it. We will run the cheap null the paper doesn't: shadow a smoothed scorer, a 4-week EMA of the score, ranked through identical machinery against the frozen scorer. After 26 live weeks it is killed unless turnover falls at least 25% and matched-formation median 20-day net excess is no worse than the frozen scorer's. If plain smoothing captures this, there's no reason to build attribution machinery.

Promising experiments (ranked)

  1. Injected-truth audit of our own promotion gates (item 2). This is the cheapest and highest-value check on the list. We have never measured gate recall, and the absolute-versus-excess bug is exactly the kind that would hide in plain sight. It costs one script and synthetic injections into placebo books.
  2. Power calculation for the live shadow head-to-head (item 5). One afternoon. It tells us whether the current promotion plan can reach a decision at all, which should be settled before more live weeks go by.
  3. Sealed holdout for shadow books from 26 October (item 1). No code, just discipline. It protects whatever result the live record eventually produces from the adaptive-narrative pattern seen in August.

Watching / not yet credible

Updates

None. This is the first edition.


This letter is drafted with AI research assistance and reviewed and signed by Bob Katz. It is research commentary, not investment advice. No result described here is a claim net of trading costs unless stated, and the experiments listed are planned tests of our own models, not findings.